january 2020 no: 1241 warwick economics …...warwick economics research papers issn 2059-4283...
TRANSCRIPT
Warwick Economics Research Papers
ISSN 2059-4283 (online)
ISSN 0083-7350 (print)
The challenges of universal health insurance in developing countries: Evidence from a large-scale randomized experiment in Indonesia
Abhijit Banerjee, Amy Finkelstein, Rema Hanna,
Benjamin Olken & Arianna Ornaghi, Sudarno Sumarto
(This paper also appears as CAGE Discussion paper 454)
January 2020 No: 1241
The challenges of universal health insurance in developing countries: Evidence from a large-scale randomized experiment in Indonesia*
Abhijit Banerjee, MIT Amy Finkelstein, MIT
Rema Hanna, Harvard University Benjamin Olken, MIT
Arianna Ornaghi, University of Warwick Sudarno Sumarto, TNP2K and SMERU
August 2019
Abstract
To assess ways to achieve widespread health insurance coverage with financial solvency in developing countries, we designed a randomized experiment involving almost 6,000 households in Indonesia who are subject to a nationally mandated government health insurance program. We assessed several interventions that simple theory and prior evidence suggest could increase coverage and reduce adverse selection: substantial temporary price subsidies (which had to be activated within a limited time window and lasted for only a year), assisted registration, and information. Both temporary subsidies and assisted registration increased initial enrollment. Temporary subsidies attracted lower-cost enrollees, in part by eliminating the practice observed in the no subsidy group of strategically timing coverage for a few months during health emergencies. As a result, while subsidies were in effect, they increased coverage more than eightfold, at no higher unit cost; even after the subsidies ended, coverage remained twice as high, again at no higher unit cost. However, the most intensive (and effective) intervention – assisted registration and a full one-year subsidy – resulted in only a 30 percent initial enrollment rate, underscoring the challenges to achieving widespread coverage.
* We thank our partners at BPJS Kesehatan, Bappenas, TNP2K, and KSP for their support and assistance. In particular, we wish to thank from BPJS Kesehatan Fachmi Idris, Mundiharno, Tono Rustiano, Dwi Martiningsih, Andi Afdal, Citra Jaya, Togar Siallagan, Tati Haryati Denawati, Atmiroseva, Muh. Syahrul, Golda Kurniawati, Jaffarus Sodiq, Norrista Ulil, and the many staff at regional BPJS offices who provided assistance; Maliki and Vivi Yulaswati from Bappenas, Jurist Tan from KSP, and Bambang Widianto and Prastuti (Becky) Soewondo from TNP2K. We thank the outstanding JPAL SEA team members for their work on this study, in particular Ignasius Hasim, Masyhur Hilmy, Amri Ilmma, Ivan Mahardika, Lina Marliani, Patrya Pratama, Hector Salazar Salame, Reksa Samudra, Nurul Wakhidah, and Poppy Widyasari. Yuanita Christayanie provided excellent research assistance. We thank SurveyMeter for outstanding data collection and fieldwork, especially Bondan Sikoki and Nasirudin. Funding from the Australian Department of Foreign Affairs and Trade and KOICA is gratefully acknowledged. The views expressed here are those of the authors, and do not necessarily reflect those of the many individuals or organizations acknowledged here.
1
I. INTRODUCTION
As developing countries emerge from extreme poverty and enter middle-income status, many aim to expand
their government-run social safety net systems (Chetty and Looney 2006). An important part of this process
is the creation of universal health insurance policies, which have expanded to many lower- and middle-
income countries over the past decade (Lagomarsino et al. 2012). In expanding health insurance, however,
emerging countries may face particularly vexing versions of the challenges faced by many developed
countries, because of the large informal sector operating outside the tax net (Jensen 2019). Some countries,
such as Thailand, have sought a single-payer-like system funded entirely out of tax revenues and
supplemented by small co-pays at the time of service (see Gruber, Hendren, and Townsend 2014), which
has been shown to improve health, but faces substantial funding challenges. Many other countries, such as
Ghana, Kenya, the Philippines and Vietnam – as well as Indonesia, which is the focus of our study – have
sought to create a contributory system with an individual mandate to reduce the financial burden on the
government. In these systems, the very poor are subsidized by tax revenues, but everyone else is required
to pay a premium, collected through a payroll tax for formal sector workers and collected directly from
individuals for everyone else.
The challenge with contributory systems, however, is that enforcing the insurance mandate for
those who must pay premiums directly is difficult. While the political and administrative challenges of
enforcing mandates are not unique to developing countries – for example, the 2010 Obamacare mandate
did not achieve universal coverage in the United States (Berchick 2018) – they are particularly difficult for
them, again because the majority of their citizens are outside the tax net. This means that the types of
penalties for non-compliance used initially in the United States under Obamacare – fines collected through
the personal income tax system – are not an option. Since developing countries have shown little appetite
for enforcing the few possible remaining sanctions on this population (e.g., denying delinquent households
the ability to enroll their kids in school), perhaps rightly, what they are left with is a toothless mandate.
In theory, a toothless mandate can create two related challenges for governments that are trying to
achieve universal or near-universal coverage: 1) low program enrollment and 2) adverse selection, where
the least healthy are more likely to enroll, raising program costs above the population average (Akerlof
1970; Einav and Finkelstein 2011). In practice, like other nations that have experimented with these
policies, Indonesia has experienced both: despite the fact that mandatory, universal health insurance was
launched in 2014, the contributory portion of the program, known as JKN Mandiri, had enrolled only 20
percent of the targeted population a year after its introduction, and its claims exceeded premiums by a ratio
2
of 6.45 to 1.1 These facts motivate the question of whether and how developing country governments can
design supplemental policies to mitigate these challenges—to boost national health insurance enrollment,
while also reining in the financial costs to the tax-funded government budget—in contexts with mandatory,
but weakly enforced, contributory health insurance programs. The aim is not necessarily for the government
to break even—it is clear that some subsidies may be needed in order to make sure that there is enough
social protection against health shocks—but to limit government spending while insuring as many people
as possible.
With this perspective in mind, in 2015, in cooperation with the Indonesian government, we
designed a large-scale, multi-arm experiment—involving almost 6,000 households—to assess three
interventions that simple economic theory suggested could increase enrollment and reduce adverse
selection in the nationally mandated insurance program. First, we examined the role of large, temporary
subsidies: we randomized households to receive subsidies of either 50 percent (“half subsidy”) or 100
percent (“full subsidy”) for the first year of enrollment. To be eligible for the subsidy, households had to
enroll within two weeks after they were offered it, akin to governments offering a large, time-limited
registration incentive. Second, we examined the role of transaction costs by randomly offering some
households at-home assistance with the online registration system, rather than traveling to a far-off
insurance office to enroll. Third, we tested for information constraints by randomly advertising three
different types of basic insurance information: the financial costs of a health episode and how they relate
to insurance prices, the two-week waiting period from enrollment to coverage (so that one could not wait
to get sick to sign up), and the fact that insurance coverage is legally mandatory.
To assess the impacts of these interventions, we utilize a number of new data sources. Notably, we
have access to the government’s administrative insurance data for program enrollees for up to 32 months
after the intervention, including detailed data on registration, premiums paid, and all claims made. We first
use these data to examine the impact of the interventions on enrollment, which we define as completing the
initial registration process. However, as the decision to stay enrolled is a dynamic one, in which households
need to pay a monthly premium, we also examine the impact of the interventions on insurance coverage,
which we define as having paid the premium for a given month and thus having insurance coverage in that
month. Next, given the extensive and detailed administrative data on claims, we examine questions relating
both to adverse selection and to the ultimate government costs per household insured under the various
policy treatments. Finally, we supplement these administrative data with a short baseline assessment survey
in which we collected data on demographics and self-reported health status prior to the intervention. Among
1 Enrollment rates are from authors’ calculation based on official membership numbers and the national sample survey, SUSENAS 2015 (BPS 2015). Claims to premium ratios are from LPEM-UI (2015).
3
other things, this baseline survey allows us to measure pre-intervention “health status” for all study
participants, regardless of whether they subsequently enrolled in the insurance program.
In the context of a toothless mandate, our findings reveal both opportunities and challenges for
increasing coverage in contributory health insurance programs in developing countries. On the one hand,
we find that temporary subsidies and assisted registration can both increase enrollment. Moreover,
temporary subsidies attract a much lower-cost population, enabling substantial increases in coverage at no
higher cost per covered unit. This increased coverage persists (albeit at a lower rate) after the subsidies end.
On the other hand, even our most intensive (and effective) intervention – assisted registration and a full
one-year subsidy – resulted in only a 30 percent initial enrollment rate; this was a substantial increase
relative to the status quo enrollment rate of 8 percent, but still a far cry from universal coverage. Our
analysis reveals specific obstacles to achieving widespread coverage stemming from limited state capacity
to facilitate enrollment and to prevent strategic short-term coverage.
Our study explicitly builds on the literature on participation in public health insurance systems—
and in social protection programs, more broadly—not only by testing the impact of these individual policy
tools on enrollment, but also by testing the relative magnitudes of relieving different participation
constraints against one another in a common real-world context.2 Theory suggests that the constraints that
we examine—reduced prices, reduced transaction costs and increased information—could each increase
enrollment in various types of public programs, including health insurance. And, in fact, the empirical
evidence is consistent with this theory: the findings from both developed and developing settings indicate
that subsidies (e.g. Thornton et al. (2010); Asuming (2013); Fischer et al. 2018; Finkelstein, Hendren and
Shepard 2019), reductions in transaction costs (e.g. Alatas et al. 2016; Bettinger et al. 2012; Dupas et al
2016), and information (Gupta 2017; Bhargava and Manoli 2015) all have the potential to increase
participation in a variety of social insurance programs, motivating our experimental design. Importantly,
our extensive high-frequency administrative data allow us to build upon this literature, as we can precisely
study whether these different types of interventions have persistent results over time, as individuals make
dynamic, and possibly strategic, decisions over insurance coverage each month.3 This is particularly
2 This study, in particular, is related to Thornton et al. (2010), which examine the impact of whether informal workers, recruited through a health insurance registration booth in the market, are randomized to receive a subsidy for contributory insurance through Nicaragua’s social security system offices or through a microfinance organization, which could potentially have been more convenient for the informal workers. Their study finds impacts of subsidies on enrollment, and therefore on utilization, but does not study how the treatments affect the degree to which the market is adversely or advantageously selected, as we do here. 3 In the developing world, there is little known about the longer-run impacts of improving health insurance take-up, as well as selection, through interventions. One notable exception is Asuming et al. (2018), which uses survey data to assess the impact of one-time subsidies on enrollment and subsequent health behaviors in Ghana, three years post-
4
important for the temporary subsidies, if “experience” with the health care system leads households to
increase their perceived value of insurance and stay covered after the subsidies expire.4
We then go further to examine not only the impacts on overall enrollment, but also whether these
interventions affect the type of individual who enrolls—as well as remains enrolled—and thus whether it is
possible to increase enrollment of low-utilization individuals enough to reduce the per-participant cost of
insurance. In the standard textbook models in which individuals differ only in their risk type, the
interventions that we test could all potentially mitigate adverse selection, since the marginal enrollees will
be lower-cost than the average enrollees (e.g. Akerlof 1970). However, in the presence of multiple
dimensions of heterogeneity, the impact of these interventions on adverse selection is theoretically
ambiguous (Einav and Finkelstein 2011); and indeed, existing evidence from health insurance studies
indicates that while such interventions can ameliorate adverse selection (e.g. Fischer et al. 2018; Finkelsten,
Hendren, and Shepard 2019), they can also, in different contexts, exacerbate it (Asuming et al. 2018; Handel
2013). It is therefore an empirical question whether varying the different insurance constraints, holding
constant the setting, leads to a different type of enrollment, and in particular one that makes a meaningful
difference in terms of participant costs.
More specifically, our three interventions produce three distinct sets of findings. First, we find that
the one-year full subsidies significantly boosted enrollment and improved selection, leading to more people
insured at the same cost to the government, even after the subsidies expired. Those offered the full subsidy
were 20.9 percentage points (almost seven times) more likely to enroll than were those in the no-subsidy
group during the active subsidy period. This increase was not driven simply by households who would have
purchased insurance anyway (“harvesting”), but rather represents a real net increase in enrollment. While
some of these households did not elect to pay premiums upon the end of the one-year subsidy, many did;
as a result, in the year after the subsidy ended, insurance coverage in the full-subsidy group remained over
twice as high as coverage in the no-subsidy group—consistent with the idea of health insurance as an
experience good.
intervention. Our high-frequency, administrative data allow us to further unpack the dynamics of selection and show how differential retention affects our understanding of these health insurance markets. The only related paper that we know that explores these issues does so in a developed country setting, studying California’s Affordable Care Act (Diamond et al. 2018). 4 Delavallade (2017) provides evidence that a related “experience” effect could be important, by showing that randomly providing households with a free preventive health visit increased their hypothetical willingness to pay for insurance in a subsequent survey.
5
Despite the fact that more households enrolled under the full-subsidy treatment, the net cost to the
government per covered person – i.e., the difference between revenues from premiums and payments to
providers, and hence the amount that would need to be covered from the general government budget -- was
similar with and without the full subsidy. Remarkably, this was true even in the first year, when the subsidy
was active and hence when the full-subsidy group brought in essentially no revenue. This is because the
subsidies brought in substantially lower-cost enrollees: relative to enrollees in the no-subsidy group, those
in the full-subsidy treatment reported better health at baseline and had fewer claims (and notably, fewer
claims for chronic conditions) during their first year of enrollment. This cost difference may also, in part,
reflect strategic timing decisions by the no-subsidy group, rather than fixed health differences alone: in fact,
the no-subsidy enrollees submitted more claims than did full-subsidy enrollees in the first three months
after enrollment, after which the difference between the two groups attenuates. Also, many enrollees in the
no-subsidy group subsequently dropped coverage – i.e., stopped paying premiums – after a few months.
Such strategic enrollment timing was less of an option for full-subsidy enrollees because the subsidy offer
was time-limited and, once enrolled, they stayed covered for the full first year. In the second year, once the
full-subsidy group had to begin paying premiums, the full-subsidy group brought in slightly more revenue
to the government since more people were enrolled (due to the experience effect highlighted above), but
the value of their claims appears similar to the value of those in the control group.
In contrast, the half-subsidy offer was less effective than the full subsidy. The half-subsidy
treatment enrolled fewer people than the full subsidy—the treatment effect was about half that of the full
subsidy, and did not appear large enough to generate an experience effect in the second year. Nor do we
observe a large selection effect on claims. Taken together, in the first year, despite bringing in more revenue
than the full subsidy, the half-subsidy treatment led to fewer households covered than the full subsidy at a
higher per-enrollee cost.
Second, while the subsidy treatments highlight that the monetary cost of insurance is a barrier to
enrollment, monetary cost alone cannot fully explain limited enrollment. Instead, we find that hassle costs
also appear to be a real barrier to participation, and one that we were not able to fully solve. Reducing
hassles by assisting with internet-based registration increased enrollment by 3.5 percentage points (41
percent). But, importantly, many more people attempted to enroll than were actually able to do so: in fact,
nearly as many people attempted to enroll in the assisted internet-based registration as in the full-subsidy
group. Combined, when offered both a full subsidy and assisted internet registration, nearly 60 percent of
households tried to enroll, but only about half were successfully able to do so. Households’ efforts were
substantially muted by technical and administrative challenges with the government’s online enrollment
system. While also reminiscent of the issues with Healthcare.gov in the United States, this particular
6
challenge stemmed from a problem common to many developing countries: Indonesia’s underlying state
civil registry, i.e., the data on who is in each family, is often inaccurate (Sumner and Kusumaningrum
2014), and since whole families must be enrolled at once (to help mitigate adverse selection), these
problems in the civil registry meant that people needed to visit an office to fix errors and sign up correctly.
Since imperfect civil registries are common throughout the developing world (Mikkelsen et al. 2015), these
types of challenges are likely to be encountered in other contexts as well.
We also find that those who enrolled in the assisted internet registration group stopped paying
premiums at a faster rate than those who enrolled under the status quo registration, perhaps reflecting that
those who selected in under this treatment might also be those who are easily discouraged by the hassle
costs involved in making payments each month. Not surprisingly, given their high dropout rate, we do not
observe any differences in the claims of the assisted registration group as compared to the status quo
registration group.
Third, none of the information treatments affected enrollment into the system. The fact that our
various information treatments had no impact suggests that lack of information may not be a key barrier,
although we cannot rule this out definitively. However, it does suggest that while information and nudge
campaigns are often an attractive policy option given their low cost (Thaler and Sunstein 2009), this does
not seem to be the primary constraint in this context.
Taken together, the most important take-away from our results is that large, temporary subsidies
can work. A common concern with offering a “free” trial period is that individuals may become used to
receiving insurance without paying, thus decreasing payments in the long-run. We find the opposite:
temporary registration incentives, featuring limited periods of free coverage before requiring premiums to
be paid, actually increase coverage and premiums paid in the subsequent year while reducing adverse
selection. This may be because many households in developing countries lack experience with insurance
(Aacharya et al. 2012), suggesting an important role for registration drives featuring temporary “free
subsidy” periods to give people experience with insurance as part of campaigns to increase enrollment.
Despite the fact that we find that these large temporary subsidies can substantially boost enrollment,
particularly among lower-cost enrollees, we did not find a silver bullet that would lead to universal (or even
close to universal) enrollment: even the most intensive intervention – assisted registration plus the free
insurance for a year – only resulted in a 30 percent initial enrollment rate. While this is substantially higher
than the status quo initial enrollment rate of 8 percent, it is still a far cry from universal enrollment;
moreover, many newly enrolled households further dropped coverage over time.
7
Nevertheless, our findings offer important insights for how to further improve these types of
programs on the margin. First, as already mentioned, a trial period of free insurance had significant positive
effects – increasing enrollment rates while substantially mitigating adverse selection – at no additional cost
to the government. Second, our results suggest that the dynamics of coverage decisions can exacerbate
adverse selection. Thus, a key administrative challenge lies not just in enforcing the enrollment mandate –
which was the premise for our interventions – but also in designing insurance regulations to prevent the
strategic timing of gaining and dropping coverage. Finally, as the assisted internet registration treatment
demonstrated, without substantial long-run investments in overall administration and infrastructure (e.g.,
improved identification systems and better internet connections), there will continue to be substantial
hassles that prevent universal insurance coverage.
The remainder of the paper is organized as follows. Section II presents the setting, the experimental
design, and the data used in the analysis. Section III presents the enrollment effects of the intervention as
well as its impacts on coverage over time. Section IV presents the selection effects, and discusses their
implications for government costs. The last section concludes.
II. SETTING, EXPERIMENTAL DESIGN AND DATA
A. Setting: the JKN Mandiri program
In January 2014, the Government of Indonesia launched Jaminan Kesehatan Nasional (JKN), a national,
contributory health insurance program aimed at providing universal coverage by 2019. JKN comprises
different sub-programs based on income and employment status. Non-poor informal workers, which
represent 30 percent of the country, are covered through a sub-program called JKN Mandiri. Under JKN
Mandiri, households must complete an initial registration process and then subsequently pay their monthly
premiums.5 While insurance enrollment is legally mandatory, the mandate is hard to enforce in practice,
and there are currently no penalties assessed on households who do not enroll.
Households may register for JKN Mandiri at any time of the year, either in person at the Badan
Penyelenggara Jaminan Sosial - Kesehatan (Social Security Administration for Health, or BPJS) office or
through the social security administration website. Households are required to register all nuclear family
5 Those below the poverty line (about the bottom 40 percent) receive fully subsidized insurance. Formal workers are covered jointly by employees and employers, and their contributions are withheld by the tax system.
8
members (e.g., father, mother, and children), as listed on their official Family Card (Karta Keluarga), which
is maintained in the civil registry by another ministry (Department of Home Affairs).
The per-person monthly premium for basic coverage (known as Class III) is IDR 25,500 (~$2),
corresponding to 3.5 percent of average monthly total expenditures for eligible households.6 The premium
that a household would pay to have JKN coverage for a year is lower than the reported yearly out-of-pocket
health expenditures for 12 percent of all non-poor informal households without health insurance, but this
percentage reaches 66 percent for households who had an inpatient episode in the last year, in which case
the median “savings” from having health insurance are large (IDR 231,341 per month).7
The premium can be paid at any social security administration office, ATM, or equipped
convenience store. Paying the premium by the 10th of a given month ensures coverage for that calendar
month. If no payment is made, coverage is deactivated after a one-month grace period. For coverage to
reactivate at a later date, the household must pay arrears, which are capped at a maximum of 6 months.8
After the program’s introduction, the government became concerned that individuals might only enroll in
JKN when they had a health emergency. To limit this, in September 2015, the government introduced a
two-week waiting period after enrollment, only after which households could submit an insurance claim.
An active membership provides coverage for healthcare costs incurred at public or affiliated clinics
and hospitals with no co-pays, although specific procedures (e.g., cosmetic surgery, infertility treatments,
orthodontics, etc.) are excluded. Primary care clinics are reimbursed under capitation based on the total
number of practitioners, the ratio of practitioners to beneficiaries, and operating hours. Hospitals are
reimbursed by case following a tariff system called INA-CBG (Indonesia Case Base Groups), in which
amounts are determined jointly by primary diagnosis and severity of the case.
B. Sample
We carried out this project in two large Indonesian cities: Kota Medan in North Sumatra and Kota Bandung
in West Java. We focused on an urban setting to abstract from supply-side issues that are likely to depress
6 There are three different classes that cover the same medical procedures, but offer different types of accommodations should an inpatient procedure be required. The per-person monthly premium during the period of the study was IDR 42,500 (~$3) for class II (3-5 beds per room) and IDR 59,500 (~$4.5) for class I (2-3 beds per room). Class III (more than 5 beds) is the most common insurance among our population of interest, with 72 percent of households in the control group enrolling in Class III insurance. 7 For each household, we compute what would have been the yearly JKN premium based on household size and compare with the yearly out-of-pocket expenditures reported in the survey using SUSENAS 2015 data (BPS 2015). 8 If no inpatient claims are submitted within 45 days from re-activation, there are no additional fees. Otherwise, the household has to pay a penalty equal to 2.5 percent of the treatment cost times the number of inactive months, up to a maximum of 12 months or IDR 30 million.
9
demand in rural areas. We chose Medan and Bandung because a significant fraction of their population was
uninsured.9 Moreover, selecting cities both on- and off-Java helps ensure representativeness of Indonesia’s
heterogeneity in culture and institutions (Dearden and Ravallion 1988).
Working with the government, we implemented the interventions in two sub-districts in Medan in
February 2015 and in eight sub-districts in Bandung in November and December 2015. The sub-districts
were selected from among those with the highest concentration of non-poor informal workers; within those
sub-districts, we randomly selected neighborhoods for the study.10 To identify JKN-eligible households
within the sampled areas, we targeted uninsured, informal workers by administering a rapid eligibility
survey to all listed households. We excluded households that already had at least one member covered by
health insurance and those that were officially below the poverty line (and thus qualified for free insurance).
Of the 52,584 listed households, 14.5 percent (7,629) satisfied the target population criteria.
When we matched our survey data with the government’s administrative data, we discovered that
some households were already covered by health insurance, even if they reported that they were not. This
was mostly an issue for the city of Medan, where the local government had recently expanded the set of
poor households who qualified for free insurance, but had not yet communicated this to the newly insured.
Since households with at least one insured member were not eligible for the study, we excluded those
already enrolled, resulting in a sample of 5,996 households.
C. Experimental design
Upon identifying an eligible household, we administered a short baseline survey (see below for details), at
the end of which the household was randomly assigned to three fully-crossed treatment arms affecting the
insurance price, the hassle cost of registration and the information available (see Figure 1).
1. Temporary Subsidy Treatments
Households were randomly selected to be in one of three groups: a control group, a full-subsidy group
covering the premiums for all family members for one year, and a half-subsidy group covering half of a
9 Other large cities, such as Jakarta, Surabaya and Makassar, introduced free local health insurance programs covering a large fraction of the population. Neither Bandung nor Medan had local programs of this type during the study period. 10 Using the 2010 Census, we chose sub-districts with a high fraction of non-poor informal workers. We excluded sub-districts with universities, large factories, or malls to avoid areas with a high concentration of temporary residents. We then randomly selected twelve kelurahan (urban municipal units) in the two sub-districts in Medan (out of 16 possible kelurahan) and four kelurahan in each sub-district in Bandung (out of 41 possible kelurahan). Within each kelurahan, we randomly selected the neighborhoods (rukun warga, also known as RW) to enumerate.
10
family’s premiums for a year.11 After the offer, the subsidy was valid for up to two weeks in Bandung and
two weeks in Medan; to be conservative and ensure sure we capture all households who enroll during the
subsidy period even accounting for data lags, our definition of households enrolled during the subsidy
period includes all households that enrolled within eight weeks of the offer date.
For logistical reasons, we could not pay half of each person’s premium. Instead, we implemented the
half subsidy through a “buy-one-get-one-free” scheme in which we paid the full premiums for half the
family members for one year, and the household was then required to pay for the other half.12 Households
chose which family members were subsidized. In theory, the government regulated that all immediate
household members be registered, so subsidizing half of the household members was roughly equivalent to
providing a 50 percent discount. The subsidy receipt for the subsidized members was conditional on
payment for the non-subsidized members for the first month, but unconditional thereafter in practice.
Households in the full subsidy period were not required to make any payments during the subsidy period.
2. Assisted Internet Registration treatment
Registering for JKN Mandiri usually requires traveling to the social security administration office in the
district capital. To reduce the hassle costs of registration, we offered half of the study households the
possibility of completing the registration process online at home with the assistance of the study
enumerator. The enumerators had internet-enabled laptops that they used to access the official social
security website. They then assisted the household with gathering the correct documentation, taking pictures
and filling in all of the forms on the website. Upon successful registration, the enumerators provided
information on payment procedures. If households wanted to think more about their options, wanted to
enroll but needed time to assemble the documentation, or had technical registration problems, the
enumerators returned within a few days to continue the enrollment process.
3. Information treatments
All study households received basic information about the insurance service coverage, the premiums, the
procedure for registration, etc. For randomly selected households in each city, we provided additional types
of information to test whether various forms of knowledge constrained enrollment.
11 In Medan, households with a positive subsidy offer were randomized to receive a one-week deadline, a two-week deadline, or the ability to choose either a one- or two-week deadline to enroll using the subsidy. In Bandung, we additionally offered a fourth subsidy sub-treatment in which households that enrolled, but did not submit an inpatient claim within a 12-month period were reimbursed 50 percent of the premiums that they had paid. Since these sub-treatments only took place in one of the two cities, we exclude them from the main analysis, but we discuss these findings below and show the results in the accompanying appendix. 12 If a family had an odd number of members, we randomly assigned the household to receive a subsidy for 𝑦 1 /2 or 𝑦 1 /2 members with equal probability. If there was only one member, the member received a full subsidy.
11
In Medan, we randomly assigned a group of households to receive additional information on the
financial costs of a health episode (“extra information treatment”). Using a script and an accompanying
booklet, we detailed the average out-of-pocket expenditures for Indonesia’s most common chronic health
conditions, as well as the cost of having a heart attack.
In Bandung, all households received basic insurance information, as well as a discussion of the out-
of-pocket expenditures associated with accessing care. However, based on discussions with the
government, we then randomly assigned households to the following two treatments: 1) a “waiting period”
treatment, in which we informed households about the new two-week waiting period between enrollment
and the start of coverage, and 2) a “mandate penalties” treatment, in which we reminded households that
enrollment is mandatory, and that there was a possibility that the government would soon introduce
regulations requiring proof of insurance to be able to renew government documents, such as passport,
driver’s license, etc.
D. Randomization design and timing
Figure 1 shows the experimental design for the cities of Medan and Bandung separately.13 Figure 2 provides
the experimental timeline. The study occurred in February 2015 in Medan and in November and December
2015 in Bandung. Subsidies were administered for twelve months after the offer for those who enrolled
within two weeks of the offer.
E. Data and Variable Definitions
We compiled two new datasets for this project. First, we conducted a short baseline survey in conjunction
with an independent and established survey firm (SurveyMeter). We administered the baseline survey
immediately following the listing questionnaire to determine eligibility. The baseline survey collected
information on the demographic characteristics of family members, self-reported health and previous health
care utilization, and existing knowledge of the program.14 Self-reported health was measured on a four-
point scale from 1 (unhealthy) to 4 (very healthy); we analyze average self-reported health across household
13 The number of households differs in each treatment for two reasons. First, while in Medan we maximized power to detect differences in enrollment, in Bandung we maximized power to detect differences in claims conditional on take-up. Since we expected greater take-up with a larger subsidy, we randomized more households into groups with smaller subsidy amounts. Second, a coding error meant that while the overall treatment probabilities were as assigned, some combinations of treatments were more likely to be randomly assigned to households than others (this coding error was corrected partway through the Bandung experiment). We include in the analysis a dummy for whether the old or new randomization was used, and reweight observations to obtain the intended cross-randomization weights so that each main treatment group has the same mix of each crossed additional treatment. 14 To minimize priming, the questions related to knowledge of the program were asked after the information on health status. The consent form only mentioned SurveyMeter and Indonesia’s National Development Planning Agency (Bappenas), the other partner in the study, but not the social security administration or JKN.
12
members. The survey was identical in Bandung and Medan, with the one exception being that we added
questions on income and employment in Bandung.
Second, we use uniquely detailed government administrative data from February 2015 to August
2018 to measure enrollment outcomes, coverage, and health care utilization.15 We track all participants for
32 months after the baseline survey. We matched the study participants to the administrative data using
individuals’ unique national identification number (Nomor Induk Kependudukan or NIK).16
We define enrollment to be the household’s successful completion of the registration process for
the national insurance program. Since a household may enroll but not actually pay any premiums, we then
also define coverage in a given month to mean that the enrolled household’s premiums were paid that
month. We use the administrative data on registration date to measure enrollment. We use the administrative
premium payment data, which report the date and value of each payment, to measure coverage.
To measure health care utilization, we analyze administrative data on all claims that are covered by
JKN in both hospitals and clinics. The hospital claim data report start and end date, diagnosis,
reimbursement value, and facility where the claim was made.17 We are able to distinguish between
outpatient and inpatient hospital claims. In contrast, all clinic claims are for outpatient procedures. The
clinic claims data report similar information to the hospital claims data, except that – due to capitation –
claim values are not available. In addition to overall claims, we report two other types of information. First,
since claims data are often noisy, we also examine the number of days until the first claim was submitted.
Second, we use the diagnoses to code whether the claim was for a chronic condition.18
15 The administrative data quality is good and has been improving over time, but some inconsistencies still arise. To ensure that we identify the correct individuals, we exclude matches when the year of birth reported in the baseline and that reported in the administrative database differ by more one year. When the same NIK links to two different membership numbers, we consider both observations as a match. When two different NIKs link to the same membership number, we exclude the observation. When enrollment date or membership type changes in subsequent extracts, we retain the information as reported in the first extract in which the individual appears. 16 About 23 percent of the individuals surveyed did not have a NIK at baseline and cannot be matched to the administrative data. We show in Column 1 of Appendix Table 1 that the probability that a household reports the NIK of at least one of its members is not differential across treatment. Given that a NIK is a requirement of enrollment, those without a NIK are likely not enrolled in JKN. 17A claim corresponds to an outpatient or inpatient event. Each event is associated with a series of diagnoses. The hospital is reimbursed for the amount that corresponds to the primary diagnosis according to the INA-CBG tariff. All exams and treatment needed for an event get reimbursed under the same claim. 18 We build our chronic classification from the Chronic Condition Indicator for the International Classification of Diseases from the Healthcare Cost and Utilization Project. This database provides information on whether diagnoses included in the ICD-10-CM: 2018 can be classified as chronic conditions. We link conditions in the ICD-10-CM: 2018 to conditions in the ICD-10: 2008 – the classification system followed by BPJS – using the first three digits of the diagnosis code. This is the lowest classification that straightforwardly corresponds across the two systems. We consider a diagnosis as chronic if it belongs to a three-digit code group with more than 75 percent chronic diagnoses.
13
F. Balance
Appendix Table 1 provides a check on the randomization by regressing various household characteristics
measured in the baseline survey on treatment dummies. Only 6 out of the 54 coefficients are significantly
different from zero at the 10 percent level, in line with what we would expect by chance.
III. IMPACTS ON ENROLLMENT AND ON SUBSEQUENT COVERAGE
A. Enrollment
Table 1 and Table 2 examine the impacts of the various treatments on enrollment – i.e., successfully
completing the registration process. We measure enrollment over the first year after the intervention date –
i.e., the date the baseline survey occurred. We estimate the following regression:
𝑦 𝛽 𝛽 𝐻𝐴𝐿𝐹 𝑆𝑈𝐵𝑆𝐼𝐷𝑌 𝛽 𝐹𝑈𝐿𝐿 𝑆𝑈𝐵𝑆𝐼𝐷𝑌 𝛽 𝐼𝑁𝑇𝐸𝑅𝑁𝐸𝑇 𝐼𝑁𝐹𝑂 ′𝛽 𝑋 ′𝛿 𝜀 (1)
where 𝐻𝐴𝐿𝐹 𝑆𝑈𝐵𝑆𝐼𝐷𝑌 , 𝐹𝑈𝐿𝐿 𝑆𝑈𝐵𝑆𝐼𝐷𝑌 and 𝐼𝑁𝑇𝐸𝑅𝑁𝐸𝑇 are dummy variables equal to 1 if household
𝑖 was randomly assigned to the respective treatment, and 𝐼𝑁𝐹𝑂 is a vector of dummies equal to 1 if
household i was randomly assigned to a particular information intervention. 𝑋 is a matrix of household-
level controls that includes dummy variables for the assignment to the other treatments (see footnote 11), a
dummy for the randomization procedure (see footnote 13) and a dummy variable for city of residence.
Regressions are weighted to reflect the desired cross-treatment randomization design (see footnote 13).
Given the household-level randomization, we report robust standard errors.19
Table 1 presents present the coefficients for 𝐻𝐴𝐿𝐹 𝑆𝑈𝐵𝑆𝐼𝐷𝑌 , 𝐹𝑈𝐿𝐿 𝑆𝑈𝐵𝑆𝐼𝐷𝑌 , and 𝐼𝑁𝑇𝐸𝑅𝑁𝐸𝑇
from equation (1), as well as the p-values from a test that the half and the full subsidy have the same
treatment effect (i.e., 𝛽 𝛽 ) and from a test that the full subsidy and the assisted internet registration
have the same effect (i.e., 𝛽 𝛽 ). Column 1 examines whether the household was enrolled within the 12
months while the subsidies were active. In Column 2, we examine whether households initiated the
enrollment process, regardless of whether they successfully enrolled.20 In Column 3, we examine
19 Note that to facilitate comparisons, we separate out interventions reported in tables. Nevertheless, the full set of indicator variables is always included. 20 For households assigned to the assisted internet registration treatment, we set attempted enrollment equal to 1 if they stated that they wanted to enroll during the visits. For households assigned to follow the status quo registration procedures, we recorded whether they showed up to the office, regardless of whether they were successful in enrolling. Since only households with a voucher had to contact the study assistant at the social security office, we do not know whether households assigned to the no-subsidy group attempted to enroll if they were not ultimately successful in enrolling. For these households, attempted enrollment is set equal to actual enrollment, a choice justified by the fact that the failure rate for households assigned to the status quo registration in the subsidy treatments was negligible.
14
enrollment within eight weeks of offer date (i.e., when the subsidy offer was valid plus some margin for
error).21 In Column 4, we consider enrollment after the subsidy offer expired, but throughout the subsidy
period (up to 1 year from the offer date).
We focus first on Panel A. Column 1 shows that the subsidies substantially increased the probability
of enrollment during the following 12 months after the offer date, while assisted registration had a positive
but smaller impact. Only about 9 percent of the no-subsidy group enrolled within the 12-month period.
Relative to this, offering the full subsidy increased enrollment by 18.6 percentage points (216 percent),
while offering the half subsidy increased enrollment by 10 percentage points (116 percent). In contrast, the
assisted internet registration treatment only increased enrollment by 3.5 percentage points (40 percent).22
The enrollment measure by itself masks the fact that many more households, particularly those in
the assisted internet treatment, attempted to enroll than were successful. Assisted internet registration led
to a 23.8 percentage point increase in attempted enrollment during the first eight weeks (column 2), but
only a 4.3 percentage point increase in successful enrollment during that period (column 3). This indicates
that less than one-fifth of the households induced by the registration assistance to attempt enrollment were
actually successful in doing so. The most common reason for unsuccessful enrollment was an inaccurate
Family Card, the official identification document (see Appendix Table 3). In order to combat adverse
selection, the government required that households enroll all nuclear family members as listed in this
document, which was pulled automatically from the digital records held by the Home Affairs Ministry. This
turned out to be problematic if the family composition had changed, but the document was not updated. In
practice, updating the card is challenging – it cannot be updated online, and instead requires at least one
trip to a Home Affairs-linked administrative office, and can often incur delays and other additional costs.
During in-person enrollment, social security administration officials use discretion to overrule the system
for cause (e.g., if households had documentation that the Home Affairs record was inaccurate), but the lack
of flexibility in the online system made web enrollment nearly impossible for many.
The evidence in Column 3 of impacts on enrollment within the first eight weeks also raises the
question of whether the interventions merely shifted forward in time an enrollment decision that would
have occurred anyway (so-called “harvesting”). This seems particularly plausible given that both the offer
of registration assistance and the subsidy offers were time-limited. Therefore, Column 4 shows the
probability of enrolling after the subsidy offer expired – specifically, after eight weeks post-offer but within
21 For all groups (including the control group), the offer date is that of the baseline survey. For subsidy group households, we consider households who have a signup date in the administrative data within eight weeks from the offer date as having enrolled using the subsidy to allow for potential delays in the data. 22 Appendix Tables 2a and 2b replicate Table 1, but disaggregate the data by city. Overall, subsidies had similar effects on actual enrollment in the two cities.
15
one year of the offer date. The results indicate that the subsidy interventions reduced the probability of
enrolling in this period, but the decline is significantly smaller than the increase due to the subsidies in the
initial period (shown in Column 3). Thus, harvesting is relatively small, accounting for no more than about
10 percent of the total additional enrollment we observed in the first eight weeks.
B. Barriers to universal enrollment
The results in Table 1, Panel A, show enrollment impacts from the intervention, but also indicate that even
with a full subsidy for a year, most people do not enroll. One explanation is that the hassle costs of
enrollment discussed above are large enough to provide a barrier, even when the insurance has no monetary
costs. To investigate this, Panel B of Table 1 shows estimates from an enhanced version of equation (1)
that also includes a full set of interactions between the (cross-randomized) subsidy treatments and the
assisted intervention treatment. Column 1 shows that even with a full subsidy and assisted internet
registration, enrollment only reached 30 percent. Column 2 shows that less than 60 percent of households
even tried to enroll when offered both free insurance for the year and assistance with registration. This
suggests that while hassle costs provide a significant barrier—even when the insurance is free—they do not
fully explain why people do not enroll.
We therefore explored other potential barriers to enrollment. Table 2 examines the role of potential
information barriers. We report the results separately by city because we tested different information
treatments in different cities, providing detailed information on heart attack costs in Medan (Panel A) and
about the nature of insurance (i.e., that enrollment is mandatory and that households must enroll in advance
of a health shock) in Bandung (Panel B). We find no statistically significant effect of any of these
information treatments. We can rule out effect sizes respectively bigger than 8.5 percentage points
(information on heart attack costs), 2.5 percentage points (information on mandates) and 3.2 percentage
points (information on waiting period).23
C. Coverage Dynamics
Insurance coverage is not a one-time decision. After the initial decision to enroll that we examined in Table
1, households must decide whether to continue to pay their monthly premiums to remain covered at any
given point in time. We now turn to the administrative data on premium payments to examine these monthly
payment decisions.
23 In Medan, we also experimentally tested whether individuals would want the offer, but procrastinate on it. Specifically, households with a positive subsidy offer were cross-randomized into different deadlines: one-week, two-week, or the possibility to choose between a one- and a two-week deadline to enroll using the subsidy. As shown in Appendix Table 4, this treatment also had effects that were indistinguishable from zero.
16
Figure 3 plots coverage by month since the offer date, by subsidy group. Coverage for a household
is defined as the premium having been paid in full for all its members that month. Payment may be made
either independently by the household or by the study; thus all households in the full subsidy group who
successfully enroll are covered for twelve months.
In the no-subsidy group, coverage slowly increased over time from 0.61 percent in the first month
of the experiment to 6.66 percent almost two years later. However, many enrollees quickly dropped
coverage; one-quarter of enrolled control group households had stopped paying their premiums three
months post-enrollment, and nearly half of the enrollees in the no-subsidy group had stopped paying their
premiums a year post-enrollment (Appendix Figure 1). The steady increase in coverage for the no-subsidy
group in Figure 3 implies that the rate of new enrollment was large enough so that net coverage rates
continued to increase despite the dropout effect.24
Interestingly, the different subsidy groups exhibited quite different levels and patterns of coverage,
both before and after the subsidies expired. In the full-subsidy group, roughly 25 percent of those offered
the full subsidy enrolled in the first two months after the offer, and their coverage mechanically remained
constant during the first year, when the subsidies were active.25 While the full-subsidy group also had a
high dropout rate after the subsidy ends (at about month 13-14), their coverage levels continued to remain
higher than the no-subsidy group, even at 20 months after the offer date.26 The fact that those brought in
with the temporary full subsidy stayed enrolled in the second year suggests a strong “experience effect,”
i.e., that these individuals may not have understood the benefits of insurance until they experienced it. This
implies that temporary subsidies can help boost enrollment past their expiration date, and may be an
important tool in boosting insurance coverage in low-enrollment settings.
As one may expect from theory, results for the half-subsidy group are somewhere between the no-
subsidy and full-subsidy results. Their coverage rate in the first year was higher than the no-subsidy group,
but far below the full-subsidy group. They also experienced a drop in coverage when the subsidy ended,
and while their coverage level was roughly flat in the second year, the no-subsidy group slowly caught up
to them. By the 20-month mark, their coverage rates appear similar.
24 The steady increase in enrollment of the no-subsidy group throughout the study period is in line with the number of enrollees going from approximately 10 million in January 2015 to more than 15 million in January 2016. 25 The slight increase in coverage shown in Figure 3 for the full-subsidy group during months 4-12 comes from the fact that a small number of households in this group enrolled after the subsidy period was over. 26 Appendix Figure 1 shows the coverage rate for the sample of those who enrolled in the first year, by month of enrollment. One can see a continuous decline in payments for those in the no- and half-subsidy groups. In contrast, there is a sharp decline for those in the full-subsidy group at month 13, the exact time when households had to start paying premiums.
17
Table 3 summarizes the coverage patterns in Figure 3.27 In Column 1, we report the percentage of
households that enrolled and had coverage for at least one month in the first year after the offer. Columns
2 and 3 decompose those with coverage in Column 1 into those who no longer had coverage by month 15
(“the dropouts”) and those who did (“the stayers”); Column 4 provides the p-value of the difference in the
dropout vs. stayer shares. Column 5 reports the percentage who had coverage in month 15, after the
subsidies ended, relative to all households in the sample; note that the interpretation in this column differs
from Column 3 since we do not condition on the household having enrolled within 1 year of the offer date.
Column 6 reports p-values for tests of whether coverage rates were the same during the subsidy period
(Column 1) and at 15 months (Column 5). Finally, Columns 7 and 8 report the same information for month
20 since offer date. In the final 3 rows of the table, we provide the p-values for tests of whether the full-
and half-subsidy coverage rates each differ from the no-subsidy coverage rates (𝛽 0 and 𝛽 0), as
well as whether the assisted-registration coverage rates differs from the status quo registration (𝛽 0).
Appendix Table 5 provides the underlying regression estimates for the p-values reported in this table.
Table 3 quantifies the magnitude of several important patterns observed in Figure 3. First, the full-
subsidy group retained substantially higher coverage than the no-subsidy group, even after the subsidies
were over. Those offered the full subsidy were 4.6 percentage points (86 percent; p-value < 0.001) more
likely than the no-subsidy group to have coverage at month 15 (column 5), and 3.9 percentage points (58
percent; p-value 0.001) more likely than the no-subsidy group to have coverage at month 20. This again
suggests that health insurance is an experience good – those who were covered for free for a limited time
were much more likely to pay for coverage afterwards than those who were never offered free insurance.
Second, despite the experience effect, we still document statistically significant declines in
coverage in the subsidy treatments. As shown in Figure 3, we observe significantly higher coverage rates
for both the full-subsidy and half-subsidy group in the first year, when the subsidy was still active (Column
1). By 20 months, after all subsidies had expired, coverage had fallen substantially and the coverage rates
at 20 months were no longer statistically distinguishable between the half-subsidy and no-subsidy group.
Even for the full-subsidy group, where we document the persistence of coverage above, comparing
Columns 1 and 7 shows that about 61 percent of those who ever had coverage in the first year had dropped
coverage by month 20 (10.6 percent in month 20 covered compared to 27.7 percent covered at some point
in the first year; p-value < 0.001). These results suggest that while temporary subsidies can lead to
27 We report means in each treatment group in this and subsequent tables to facilitate comparisons both across time and across treatment groups. The means for each cell are calculated using the weights described in footnote 13, so that each treatment group shown has the same (weighted) combination of sub-treatments; that is, half subsidy has the same weighted mix of status quo vs. assisted internet registration as full subsidy, and so on.
18
substantial increases in coverage even after the subsidies are over, only about 40 percent of those subsidized
continue to retain coverage.
Finally, it is important to note that while the assisted-registration group saw a slight increase in
coverage initially (Column 1), their coverage rate quickly converged to that of the control group. This
suggests that some of the households brought into the insurance system by reducing hassles may have been
particularly sensitive to the hassles of paying each month, leading to the increased dropout rate. One
possible reason is that while the assisted internet registration made registration easier, it did not resolve the
hassles of paying one’s premium, which still needed to be done at an office, ATM, or convenience store.
IV. Selection Impacts and its Implications on Government Costs
A. Impacts on Selection
Subsidies are a textbook response to concerns about adverse selection, since in standard models they will
induce lower-cost individuals to enroll. To examine the types of people who enroll under different
intervention arms, we draw on two sources of data: self-reported health from the baseline survey, and
administrative claims data among those who enrolled. While these two measures capture different objects
– namely, health and healthcare usage – perhaps not surprisingly, enrollees with better self-reported health
indeed tend to have fewer claims (see Appendix Table 6).
Table 4 shows the results.28 It shows various measures of health and healthcare use for those who
enrolled and had coverage for at least one month during the first year (i.e., as measured in column 1 of
Table 3). Column 1 indicates that the marginal household who received coverage in response to the
subsidies had a higher level of self-reported health at baseline than enrollees in the no-subsidy group. Those
enrolling with the subsidies had an average self-reported health score that is about 4.5 percent higher than
that of no-subsidy enrollees, with both subsidy treatment effects significant at the 5 percent level. The
effects of assisted internet registration were smaller, but in the same direction, and statistically significant
at the 10 percent level.29
The remaining columns of Table 4 examine healthcare usage of households who enrolled and had
coverage for at least one month during the first year. We examine all claims for the 12 months after the
enrollment date; by examining a fixed number of months since enrollment date regardless of when
28 The regressions that calculate these p-values are provided in Appendix Table 7. 29 Appendix Table 8 shows that the results holds also if self-reported health is measured in as the self-reported health of the least healthy family member. In addition, households who enrolled under the full subsidy treatment were also less likely to have a family member above 60.
19
households enrolled, we can abstract from the feature that temporary subsidies may drive households to
enroll earlier in a calendar year thus mechanically affecting length of insurance coverage. We focus on
three main indicators: whether the household had any claim (column 2), the total number of visits made
(column 6), and the total value of claims paid (column 10). We then subdivide claims into outpatient,
inpatient, and chronic. We also examine the number of days to first claim, which can provide greater
precision than the value of claims, which tend to have a large right tail (Aron-Dine et al. 2015).
Consistent with the results on self-reported health in column 1, the claims analysis in the remaining
columns also indicates that those who enrolled under the full subsidy were healthier and lower-cost.
Households in the full-subsidy group were less likely to submit claims. For example, in the no-subsidy
group, 62 percent had any claim compared to 48 percent in the full-subsidy group (column 2; p-value 0.040).
Those in the full-subsidy group were also less likely to have had a claim for a chronic, ongoing condition:
27 percentage points for the no-subsidy group compared to 17 percentage points for the full-subsidy group
(column 5; p-value 0.082). Results for the half-subsidy group are mostly qualitatively similar to the full-
subsidy group, but smaller in magnitude and never statistically significantly different from the no-subsidy
group. The same is true of the results for the assisted internet registration group.
In addition to having fewer overall claims, claims in the full subsidy group were less likely to be
“large claims” that suggest a substantial health emergency. This is shown in Figure 4, which reports the
probability distribution function of the value of inpatient claims submitted within twelve months since
enrollment, by treatment status, for those who enrolled within a year since offer date and paid for at least
one month. The distribution of values of claims for the full-subsidy group is markedly left-shifted relative
to the no-subsidy group. Again, the same is true – although less pronounced – in comparing the half-subsidy
and no-subsidy groups. The differences across groups are statistically significant according to a
Kolmogorov-Smirnov test for equality of distribution functions (p=0.012 for the test of equality between
the distribution of the half-subsidy and no-subsidy groups and p=0.001 for the test of equality between the
distribution of the full-subsidy and no-subsidy groups). In short, when they use the health care system, those
whose coverage was heavily subsidized have less expensive health incidents.
On net, the fact that the full-subsidy group had fewer claims, and that these claims were small,
results in substantial reductions in claims expenditures from the insurer. In particular, the full-subsidy group
had average claims that were 40 percent lower in value than those in the no-subsidy group (column 10 of
Table 4; p-value 0.095) and on average waits 30 percent longer before submitting their first claim (column
11; p-value 0.006).
B. Dynamics and Selection
20
An important question is whether the fact that households can time enrollment and dropout
decisions exacerbates adverse selection. We investigate both a) whether households in the no-subsidy
group, who do not face a time limited enrollment period, are more likely to time enrollment to when they
are likely to have a claim, and b) how those who choose to retain coverage differ from those who drop.
Figure 5 begins by plotting the number of claims by month since enrollment, separately by subsidy
treatment groups among households who enrolled within one year since offer and had coverage for at least
one month over that period, along with 95 percent confidence intervals. Those who enrolled without the
subsidy appear to have submitted more claims in the first few months upon enrollment than did the
households in the full-subsidy group, but over time this difference became less stark and, by the end of the
period, they displayed similar patterns in number of claims. Households in the half-subsidy group also
submitted more claims than households in the full-subsidy group, and even submitted claims for a higher
value than the no-subsidy group in a handful of months.30 Combined with the payments findings in the
previous section (i.e., Figure 3), this suggests that no-subsidy households may have had large claims once
they enrolled, but then stopped paying premiums (i.e., dropped coverage). In contrast, the subsidy groups
brought in healthier people, who kept paying premiums longer in the first year while the subsidies were
active (see Figure 3), and had smaller claims throughout the year (Figure 5).
Table 5 investigates differential selection in terms of who retained coverage. For each treatment,
we divide those who enrolled in the first year into “dropouts” – those who did not still have coverage in
month 15 – and “stayers” – those who did. The coverage rates of these two groups are shown in Table 3.
Several results are worth highlighting. In the full-subsidy group, those who retained coverage had higher
baseline self-reported health than those who did not (column 1; p-value 0.068). On the other hand, the
stayers were also more likely to have had claims (column 2; p-value 0.005) and to have had more visits
(column 6; p-value 0.002). These were particularly likely to be outpatient claims/visits and those for chronic
conditions, rather than inpatient claims. The half-subsidy group showed a similar pattern of claims.31 The
pattern for the no-subsidy group is more ambiguous, with the dropouts more likely to have had an inpatient
claim, but having had fewer overall visits.
30 Appendix Table 9 formally confirms this result. In the first 3 months that the households were enrolled in insurance (Panel A), full-subsidy households were less likely to submit inpatient or outpatient claims, had lower overall claims than the control group, and their inpatient claims were on average for lower values. The coefficient on the half-subsidy group is generally negative, but the difference is not always statistically significant. Months four through twelve after enrollment (Panel B) show that, over time, the difference disappears: all of the treatment groups display a very similar pattern of claims, although the coefficient for the full-subsidy group is overall still negative. 31 Appendix Table 10 presents the equivalent results broken down by the assisted internet registration treatment, and finds a similar pattern: stayers were more likely to have had claims, particularly inpatient and outpatient claims. Appendix Table 11 presents the regressions from which we calculate the p-value of the difference in means reported in Table 5 and in Appendix Table 10.
21
The results from the subsidy treatments continue to suggest an experience effect: those who stayed
were those who made use of the system, even for smaller outpatient or chronic conditions. They also raise
the possibility that allowing relatively small payments from a plan (as opposed to a high-deductible plan
that only covers catastrophic expenses) may be important for continuing to entice healthy people to remain
covered.
C. Implications for government costs
The selection patterns indicate that the subsidies brought in healthier enrollees, while the coverage
dynamics indicate that not only were no-subsidy enrollees sicker and higher-cost, but that they strategically
timed enrollment to coincide with major health expenditures, and were quicker to drop coverage (i.e., stop
paying premiums) after a few months. In Table 6, we examine the implications of these results for the net
costs to the government. The results indicate that the subsidies covered more households at similar cost per
covered household.32
Table 6 shows net revenues with and without accounting for capitation payments by month, which
can be decomposed into revenues from premiums and government expenditures as a result of claims.33
Revenues are defined as premiums paid by enrollees.34 Claims expenditures are defined as the value of
claims paid. In columns 2 to 5, we focus on revenues and expenditures per household-month covered; this
provides us with estimates of the additional revenue and expenditure for each additional household covered
in a given month. As an alternative way of presenting the same results, columns 6 through 9 report the
results for all households in the sample regardless of whether they enrolled in the insurance, thus providing
us the total revenues and costs of offering the policy; these estimates reflect the corresponding cells in
columns 2 through 5, scaled by the number of covered households in that group.
Panel A explores the impacts on government net costs during the period when the subsidy was in
effect. While the subsidy was active, on net the government lost around IDR 125,000 (~$9) per household-
32 Appendix Table 12 presents equivalent results split by assisted internet registration vs. status quo registration, and finds no substantial differences in net costs to the government. Appendix Table 13 reports the regressions that correspond to the p-values reported in Table 6 and Appendix Table 12. 33 Capitation payments depend on the number of enrollees that declare the facility as their primary provider, the total number of practitioners, the ratio of practitioners to beneficiaries, and operating hours, and range between IDR 3,000-6,000 per enrollee for puskesmas and IDR 8,000-10,000 for clinics. Given that approximately 80 percent of JKN Mandiri enrollees declare puskesmas and 20 percent declare clinics as their primary health facility, for these calculations we assume capitation payments to be IDR 6,800 per enrollee per month. Capitation payments are only paid to healthcare facilities in months in which the household paid the premium. 34 They should therefore be mechanically zero for the full-subsidy group while the subsidy is in effect, but are not literally zero since a few households in this group enrolled after the time period the subsidy offer was in effect and therefore had to pay premiums.
22
month covered in the no-subsidy group (Column 5, Panel A). By comparison, over this same period, on net
the government lost only about IDR 50,000 (~$3.50) per household-month covered in the full-subsidy
group (Column 5, Panel A). In other words, the net cost to the government per covered household-month
in the subsidy group was no higher than in the no-subsidy group (p-value = 0.19), even taking into account
that the government received essentially no revenue from the subsidy group. This is because the decline
in average claims between the full-subsidy and no-subsidy groups (Column 3: decline of IDR 152,000 per
covered household-month; p-value 0.026) was even larger than the forgone revenue from not collecting
premiums (Column 2: decline of IDR 71,000 per covered household-month; p-value <0.001).35 As a result,
the full subsidy resulted in over eight times more covered household-months (Column 1), at no higher cost
to the government per household-month covered (Column 5).
Of course, there are more people covered, so this policy does entail an increase in the total amount
spent by the government in total. Looking over the entire sample (i.e., not conditioning on enrollment),
Column 9 indicates that in the no-subsidy group the government cost was IDR 3,000 (~$0.20) per eligible
household per month, while with the full-subsidy the government cost was IDR 6,000 (~$0.40) per eligible
household per month.
Panel B explores what happened in the year after the subsidies were withdrawn. As shown in Table
3, there was some persistent increase in coverage in the full-subsidy group. However Table 5 shows that
despite being healthier initially, households in the full-subsidy group that retained coverage (i.e., paid
premiums) after the subsidy ended were more likely to have had a claim during the first year than those
who dropped coverage after the subsidy ended. As a result, those who retained coverage had similar
average claims after the subsidy ended to those in the no-subsidy group. On net, Table 6, Panel B (Column
5) indicates that government costs were not statistically different per covered household-month for the no-
subsidy group and the full-subsidy group in the period after the subsidy ended, though the point estimates
suggest that that the government lost about half as much per covered household-month in the full subsidy
group compared to the no-subsidy group (IDR 47,000 in net government costs per covered household-
month in the full-subsidy group; IDR 102,000 in net government costs per covered household-month in the
no-subsidy group). Thus, in the year after the temporary full subsidy ended, we estimate that twice as many
household-months were covered (Panel B, Column 1), at no higher cost per covered household-month
(Panel B, Column 5). Looking over the entire sample (i.e., not conditioning on enrollment), Column 9
35 For household-months covered in the half-subsidy group, the net losses are similar to those in the no-subsidy group (about IDR 160,000 per covered household-month); again, the fact that net revenue losses are only slightly larger for the half-subsidy group than for the no-subsidy group – despite mechanically lower revenues – reflects the healthier composition of the half-subsidy pool.
23
indicates virtually identical government expenditures (about IDR 5,000 per person) in the year after the
subsidies end for the full-subsidy group compared to the no-subsidy group.
Putting this all together, one can calculate the bottom-line implications for the government for
offering a temporary full subsidy to a given population. In the year the subsidy was in effect, the government
doubled its net budgetary contribution for this population (from IDR 3,000 to IDR 6,000 per person
offered). During that year, coverage expanded dramatically – from 6.3 percent of the population to 27.7
percent of the population. In the subsequent year, the bottom line for the government was the same – about
IDR 5,000 per person in the population – regardless of treatment. But the full-subsidy group had, on net,
58 percent more people covered, with the same total government expenditure. This is very far from
universal coverage – the full-subsidy group had 10.6 percent covered at 20 months after the project started,
compared to 6.7 percent in the no-subsidy group – but it represents meaningfully more people covered with
no additional ongoing cost to the government.
V. CONCLUSION
As incomes have risen in emerging economies, there has been a growing move to increase coverage of
social insurance programs. However, insurance mandates can be difficult to enforce. We examine the
impact of temporary insurance subsidies – which must be taken up within a month of offer and only last a
year – reduced hassle costs, and information provision on insurance coverage in a mandated insurance
setting. We find that offering a full, but temporary, subsidy was effective at increasing enrollment and
helped attract healthier enrollees. Because of the healthier selection and also the strategic dynamic
adjustment of coverage and claims in the no-subsidy group – the no-subsidy group timed its enrollment to
coincide with high expenditures and quickly dropped coverage a few months later – the net cost to the
government per covered household-month of the full subsidy is no higher, even despite the cost of the
subsidies. Importantly, subsidies induced higher enrollment even after they expired, in line with health
insurance being an experience good. As a result, after the subsidy period was over, the government was
able to cover substantially more people at a roughly similar net cost.
But, at the same time, our findings also highlight challenges that governments face when aiming to
achieving universal health coverage through a contributory system. While both subsidies and assisted
enrollment increased enrollment rates, even the most aggressive interventions – a full subsidy for a year
and internet-assisted enrollment, only led to 30 percent enrollment – a substantial increase from the 8
percent enrollment in the status quo group, but a far cry from universal enrollment. Some of this reflects
administrative challenges: almost 60 percent of households in the full subsidy, internet assisted registration
24
treatment tried to enroll – double the amount who actually did so. This underscores how weak social
insurance infrastructure (in this case, the underlying social registry) can create obstacles to universal
enrollment and suggests that long-run solutions to universal coverage are only feasible through
strengthening overall administrative structures.
References
Acharya, Arnab, Sukumar Vellakkal, Fiona Taylor, Edoardo Masset, Ambika Satija, MargartBurke, and Shah Ebrahim. 2012. “Impact of National Health Insurance for the Poor and the Infor-mal Sector in Low-and Middle-Income Countries.” Systematic Review, The EPPI-Centre .
Akerlof, George. 1970. “The Market for Lemons: Quality Uncertainty and the Market Mecha-nism.” Quarterly Journal of Economics 84 (3):488–500.
Alatas, Vivi, Abhijit Banerjee, Rema Hanna, Benjamin A. Olken, Ririn Purnamasari, and MatthewWai-Poi. 2016. “Self-targeting: Evidence from a Field Experiment in Indonesia.” Journal of PoliticalEconomy 124 (2):371–427.
Aron-Dine, Aviva, Mark Cullen, Liran Einav, and Amy Finkelstein. 2015. “Moral Hazard in HealthInsurance: Do Dynamic Incentives Matter?” Review of Economics and Statistics 97 (4):725–741.
Asuming, Patrick O. 2013. “Getting the Poor to Enroll in Health Insurance, and its effects on theirhealth: Evidence from a Field Experiment in Ghana.” Working Paper.
Asuming, Patrick Opoku, Hyuncheol Bryant Kim, and Armand Sim. 2019. “Long-Run Conse-quences of Health Insurance Promotion when Mandates are not Enforceable: Evidence from aField Experiment in Ghana.” Working Paper.
Badan Pusat Statistik. 2015. “Survei Sosial Ekonomi Nasional Maret (KOR).”
Berchick, Edward R, Emily Hood, and Jessica C Barnett. 2018. “Health Insurance Coverage in theUnited States: 2017.” Current Population Reports, P60-264, US Government Printing Office, Washing-ton, DC .
Bettinger, Eric P, Bridget Terry Long, Philip Oreopoulos, and Lisa Sanbonmatsu. 2012. “The Role ofApplication Assistance and Information in College Decisions: Results from the H&R Block FAFSAExperiment.” The Quarterly Journal of Economics 127 (3):1205–1242.
Bhargava, Saurabh and Dayanand Manoli. 2015. “Psychological Frictions and the IncompleteTake-up of Social Benefits: Evidence from an IRS Field Experiment.” American Economic Review105 (11):3489–3529.
Capuno, Joseph J, Aleli D Kraft, Stella Quimbo, Carlos R Tan, and Adam Wagstaff. 2016. “Effectsof Price, Information, and Transactions Cost Interventions to Raise Voluntary Enrollment in aSocial Health Insurance Scheme: a Randomized Experiment in the Philippines.” Health Economics25 (6):650–662.
Chetty, Raj and Adam Looney. 2006. “Consumption Smoothing and the Welfare Consequences ofSocial Insurance in Developing Economies.” Journal of Public Economics 90 (12):2351–2356.
25
Dearden, Lorraine and Martin Ravallion. 1988. “Social Security in a Moral Economy: An EmpiricalAnalysis for Java.” The Review of Economics and Statistics :36–44.
Delavallade, Clara. 2017. “Quality Health Care and Willingness to Pay for Health Insurance Re-tention: A Randomized Experiment in Kolkata Slums.” Health Economics 26 (5):619–638.
Diamond, Rebecca, Michael J Dickstein, Timothy McQuade, and Petra Persson. 2018. “Take-Up,Drop-Out, and Spending in ACA Marketplaces.” NBER Working Paper No. 24668.
Einav, Liran and Amy Finkelstein. 2011. “Selection in Insurance Markets: Theory and Empirics inPictures.” Journal of Economic Perspectives 25 (1):115–38.
Finkelstein, Amy, Nathaniel Hendren, and Mark Shepard. 2019. “Subsidizing Health Insurancefor Low-Income Adults: Evidence from Massachusetts.” American Economic Review .
Fischer, Torben, Markus Frölich, and Andreas Landmann. 2018. “Adverse Selection in Low-Income Health Insurance Markets: Evidence from a RCT in Pakistan.” IZA DP No. 11751.
Fitzpatrick, Anne and Rebecca Thornton. 2018. “The Effects of Health Insurance within Families:Experimental Evidence from Nicaragua.” World Bank Economic Review .
Gerard, François and Gustavo Gonzaga. 2018. “Informal Labor and the Efficiency Cost of So-cial Programs: Evidence from the Brazilian Unemployment Insurance Program.” NBER WorkingPaper No. 22608.
Gruber, Jonathan, Nathaniel Hendren, and Robert M Townsend. 2014. “The Great Equalizer:Health Care Access and Infant Mortality in Thailand.” American Economic Journal: Applied Eco-nomics 6 (1):91–107.
Gupta, Sarika. 2017. “Perils of the Paperwork: The Impact of Information and Application Assis-tanceon Welfare Program Take-Up in India.” Working Paper.
Handel, Benjamin R. 2013. “Adverse Selection and Inertia in Health Insurance Markets: WhenNudging Hurts.” American Economic Review 103 (7):2643–82.
Jensen, Anders. 2019. “Employment Structure and the Rise of the Modern Tax System.” NBERWorking Paper No. 25502.
Lagomarsino, Gina, Alice Garabrant, Atikah Adyas, Richard Muga, and Nathaniel Otoo. 2012.“Moving Towards Universal Health Coverage: Health Insurance Reforms in Nine DevelopingCountries in Africa and Asia.” The Lancet 380 (9845):933–943.
Mikkelsen, Lene, David E Phillips, Carla AbouZahr, Philip W Setel, Don De Savigny, RafaelLozano, and Alan D Lopez. 2015. “A Global Assessment of Civil Registration and Vital Statis-tics Systems: Monitoring Data Quality and Progress.” The Lancet 386 (10001):1395–1406.
26
Sumner, Cate and Santi Kusumaningrum. 2014. AIPJ Baseline Study on Legal Identity: Indonesia’sMissing Millions. Australia Indonesia Partnership for Justice.
Thaler, Richard H and Cass R Sunstein. 2009. Nudge: Improving Decisions about Health, Wealth, andHappiness. Penguin.
Thornton, Rebecca L, Laurel E Hatt, Erica M Field, Mursaleena Islam, Freddy Solís Diaz, andMartha Azucena González. 2010. “Social Security Health Insurance for the Informal Sector inNicaragua: a Randomized Evaluation.” Health Economics 19 (S1):181–206.
Wagstaff, Adam, Ha Thi Hong Nguyen, Huyen Dao, and Sarah Bales. 2016. “Encouraging HealthInsurance for the Informal Sector: a Cluster Randomized Experiment in Vietnam.” Health Eco-nomics 25 (6):663–674.
27
Figure 1: Experimental Design
Standard
information
Mandate
information
Standard
information
Mandate
information
Standard
information236 307 241 274
Waiting period
information232 300 297 349
Standard
information160 153 77 82
Waiting period
information141 114 100 91
Standard
information85 40 62 54
Waiting period
information63 51 70 53
Standard
information
Extra
information
Standard
information
Extra
information
No subsidy 37 63 134 237 471
Half subsidy 171 215 26 66 478
Full subsidy 176 54 170 97 497
Registration
treatment totals
No subsidy
Half subsidy
Full subsidy
Registration treatment totals
Subsidy
treatment totals
Bandung
716 730
Subsidy
treatment totals
Medan
2236
918
478
17501882
Status quo Assisted internet registration
Status quo Assisted internet registration
Standard
information
Mandate
information
Standard
information
Mandate
information
Standard
information236 307 241 274
Waiting period
information232 300 297 349
Standard
information160 153 77 82
Waiting period
information141 114 100 91
Standard
information85 40 62 54
Waiting period
information63 51 70 53
Standard
information114 86 170 111
Waiting period
information101 86 131 119
Standard
information
Extra
information
Standard
information
Extra
information
No subsidy 37 63 134 237 471
Half subsidy 171 215 26 66 478
Full subsidy 176 54 170 97 497
Registration
treatment totals
918
Subsidy
treatment totals
Bandung
716 730
Subsidy
treatment totals
Medan
2236
918
478
22812269
Status quo Assisted internet registration
Status quo Assisted internet registration
No subsidy
Half subsidy
Full subsidy
Registration treatment totals
Bonus subsidy
Note: This figure shows the randomization into each treatment arm, by city. Each cell reports the number of households allocated tothe specific treatment cell.
28
Figu
re2:
Tim
elin
e
23
45
67
89
1011
121
23
45
67
89
1011
121
23
45
67
89
1011
121
23
45
67
8
Medan
Base
line
and
trea
tmen
tSu
bsid
y pe
riod
Adm
inis
trat
ive
data
Bandung
Base
line
and
trea
tmen
tSu
bsid
y pe
riod
Adm
inis
trat
ive
data
2018
2015
2016
2017
Not
e:Th
isfig
ure
prov
ides
ati
mel
ine
byci
ty.
The
base
line
and
trea
tmen
tof
fer
occu
rred
inth
esa
me
hous
ehol
dvi
sit,
inFe
brua
ry20
15in
Med
anan
dN
ovem
ber/
Dec
embe
r20
15in
Band
ung.
Subs
idie
sw
ere
disb
urse
dov
erth
eco
urse
ofth
efo
llow
ing
twel
vem
onth
saf
ter
offe
r.W
eac
cess
edth
ead
min
istr
ativ
eda
tain
Apr
il20
15,A
pril
2016
,May
2017
,and
Aug
ust
2018
.
29
Figure 3: Insurance Coverage by Month since Offer, by Subsidy Treatment
0.1
.2.3
shar
e of
HH
s w
ith c
over
age
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20months since offer date
No Subsidy Half Subsidy Full Subsidy
Note: This figure shows mean insurance coverage by month since the offer for households assigned to different subsidy treatments,with 95% confidence intervals for the mean. Means are weighted to reflect the intended randomization. Coverage for a household isdefined as the premium having been paid in full for all its members that month. The sample size is 5996 households.
Figure 4: Distribution of Inpatient Claims, by Subsidy Treatment in Year 1 since Enrollment
0.1
.2.3
.4de
nsity
0 5 10 15value of claims (IDR 1,000,000)
No Subsidy Half Subsidy Full Subsidy
Note: This figure shows the probability distribution function of the value of inpatient claims submitted within the first twelve monthssince enrollment by subsidy treatment. The unit of observation is a single claim. The sample is restricted to households who enrolledwithin one year since offer and had coverage for at least one month over the same time period. The sample size is 3827 inpatientclaims.
30
Figure 5: Number of Claims, by Month since Enrollment and by Subsidy Treatment
0.2
.4.6
.81
num
ber o
f cla
ims
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20months since enrollment
No Subsidy Half Subsidy Full Subsidy
Note: The figure shows the mean number of claims in each month since enrollment with 95% confidence intervals for the mean. Meansare weighted to reflect the intended randomization. The sample is restricted to households who enrolled within one year since offerand had coverage for at least one month over the same time period. The sample size is 749 households.
31
Table 1: Effect of Temporary Subsidies and Assisted Internet Registration on Year 1 Enrollment
(1) (2) (3) (4)
Full subsidy 0.186*** 0.276*** 0.209*** -0.023**(0.020) (0.020) (0.018) (0.010)
Half subsidy 0.100*** 0.161*** 0.114*** -0.014*(0.014) (0.016) (0.013) (0.008)
Assisted internet registration 0.035*** 0.237*** 0.043*** -0.008(0.011) (0.011) (0.009) (0.006)
No subsidy mean 0.086 0.099 0.030 0.056
Half subsidy = full subsidy 0.000 0.000 0.000 0.442Assisted internet registration = full subsidy 0.000 0.107 0.000 0.266
0.223*** 0.557*** 0.239*** -0.016(0.025) (0.026) (0.024) (0.012)
Full subsidy and status quo registration 0.193*** 0.191*** 0.200*** -0.006(0.027) (0.025) (0.024) (0.014)
0.148*** 0.418*** 0.162*** -0.014(0.025) (0.029) (0.023) (0.012)
Half subsidy and status quo registration 0.092*** 0.099*** 0.087*** 0.005(0.016) (0.015) (0.013) (0.011)
No subsidy and assisted internet registration 0.028*** 0.179*** 0.023*** 0.005(0.011) (0.012) (0.007) (0.008)
No subsidy, status quo registration mean 0.078 0.018 0.018 0.060
Decomposition
Panel A: Main effects
Panel B: Interacted specification
Table 1
Half subsidy and assisted internet registration
Enrolled within 8 weeks of offer date
Enrolled after 8
weeks, but within 1 year of offer date
Enrolled within 1 year
Attempted to enroll within 8 weeks of offer date
P-value of test of hypothesis
Full subsidy and assisted internet registration
Note: This table shows the effect of subsidies and assisted internet registration on enrollment in year 1 since offer. The sample size is5996 households. In Panel A, we regress each outcome on indicator variables for treatment assignment, an indicator variable for therandomization procedure used and an indicator variable for the study location (equation (1)). The omitted category is no subsidy forthe subsidy treatments and status quo registration for the assisted internet registration treatment. The p-values reported are from atest of the difference between the half subsidy and full subsidy treatments (β1 = β2) and assisted internet registration and full subsidytreatments (β1 = β3). Panel B shows the effect of the interacted treatments on enrollment in year 1. The omitted category is no subsidyand status quo registration treatment. All regressions are estimated by OLS and weighted to reflect the intended cross-randomization.Robust standard errors are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
32
Table 2: Effect of Information on Year 1 Enrollment, by CityTable 2
(1) (2) (3) (4)
0.029 -0.000 0.034 -0.005(0.029) (0.030) (0.025) (0.016)
Observations 1446 1446 1446 1446No information mean 0.190 0.369 0.131 0.059
0.004 0.015 -0.001 0.005(0.011) (0.012) (0.009) (0.007)0.011 0.006 0.009 0.002
(0.011) (0.012) (0.009) (0.007)
Observations 4550 4550 4550 4550No information mean 0.123 0.188 0.078 0.045
xx
Decomposition
Enrolled after 8
weeks, but within 1 year of offer date
Attempted to enroll within 8 weeks of offer date
Enrolled within 8 weeks of offer date
Enrolled within 1 year
Information on cost of treatment for heart attack
Information on possible mandate penalties
Information on two weeks waiting period
Panel B: Bandung
Panel A: Medan
Note: This table shows the effect of the information treatments on enrollment, by city. We regress each outcome on indicator variablesfor treatment assignment, an indicator variable for the randomization procedure used and an indicator variable for the study location(equation (1)). The omitted category is no information for all information treatments. All regressions are estimated by OLS andweighted to reflect the intended randomization. Robust standard errors are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
33
Tabl
e3:
Insu
ranc
eC
over
age,
byTe
mpo
rary
Subs
idie
san
dA
ssis
ted
Inte
rnet
Reg
istr
atio
nT
able
3
Dro
pout
sS
taye
rs
Had
co
vera
ge f
or
at le
ast o
ne
mon
th
Did
not
hav
e co
vera
ge in
m
onth
15
Had
co
vera
ge in
m
onth
15
P-V
alue
(2)
vs (
3)
Had
co
vera
ge in
m
onth
15
P-V
alue
(1)
vs (
5)
Had
co
vera
ge in
m
onth
20
P-V
alue
(1)
vs (
7)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Full
sub
sidy
0.27
70.
184
0.09
30.
000
0.09
90.
000
0.10
60.
000
Hal
f su
bsid
y0.
171
0.10
70.
064
0.00
50.
074
0.00
00.
074
0.00
0N
o su
bsid
y0.
063
0.02
40.
038
0.00
70.
053
0.02
10.
067
0.44
5
Ass
iste
d in
tern
et r
egis
trat
ion
0.14
00.
082
0.05
80.
003
0.06
90.
000
0.07
50.
000
Stat
us q
uo r
egis
trat
ion
0.11
70.
059
0.05
80.
862
0.06
90.
000
0.08
30.
000
Full
sub
sidy
= n
o su
bsid
y0.
000
0.00
00.
000
0.00
00.
001
Hal
f su
bsid
y =
no
subs
idy
0.00
00.
000
0.00
30.
032
0.30
7A
ssis
ted
inte
rnet
reg
istr
atio
n =
st
atus
quo
0.02
80.
006
0.96
10.
937
0.29
8
Enr
olle
d w
ithi
n 1
year
of
offe
r da
te a
nd
P-va
lue
of te
st o
f hy
poth
esis
Not
e:Th
ista
ble
show
sm
ean
cove
rage
byte
mpo
rary
subs
idy
and
assi
sted
inte
rnet
regi
stra
tion
.A
hous
ehol
dis
cons
ider
edas
havi
ngin
sura
nce
cove
rage
ifth
epr
emiu
mw
aspa
idfo
ral
lits
mem
bers
.Th
esa
mpl
esi
zeis
5996
hous
ehol
ds.
The
p-va
lues
atth
ebo
ttom
ofth
eta
ble
are
from
regr
essi
ons
ofea
chou
tcom
eon
indi
cato
rva
riab
les
for
trea
tmen
tas
sign
men
t,an
indi
cato
rva
riab
lefo
rth
era
ndom
izat
ion
proc
edur
eus
edan
dan
indi
cato
rva
riab
lefo
rst
udy
loca
tion
(equ
atio
n(1
)).
Stan
dard
erro
rsar
ero
bust
.T
hep-
valu
esre
port
edar
efr
oma
test
ofth
edi
ffer
ence
betw
een
the
nosu
bsid
yan
dfu
llsu
bsid
ytr
eatm
ents
(β2=
0),b
etw
een
the
nosu
bsid
yan
dha
lfsu
bsid
ytr
eatm
ents
(β3=
0)an
dbe
twee
nth
est
atus
quo
and
assi
sted
inte
rnet
regi
stra
tion
trea
tmen
ts(β
4=
0).
The
p-va
lues
inco
lum
ns(4
),(6
)and
(8)a
refr
omre
gres
sion
sin
whi
chw
est
ack
the
outc
omes
bein
gco
mpa
red
and
regr
ess
them
onin
dica
tor
vari
able
sfo
rtr
eatm
ent
assi
gnm
ent,
the
inte
ract
ion
ofth
ese
indi
cato
rva
riab
les
wit
han
indi
cato
rva
riab
lefo
rth
eou
tcom
eth
eob
serv
atio
nre
fers
to,a
nin
dica
tor
vari
able
for
the
rand
omiz
atio
npr
oced
ure
used
and
anin
dica
tor
vari
able
for
stud
ylo
cati
on.I
nth
ese
regr
essi
ons
stan
dard
erro
rsar
ecl
uste
red
atth
eho
useh
old
leve
l.A
llre
gres
sion
sar
ees
tim
ated
byO
LSan
dw
eigh
ted
tore
flect
the
inte
nded
rand
omiz
atio
n.A
ppen
dix
Tabl
e5
prov
ides
the
regr
essi
ones
tim
ates
behi
ndth
enu
mbe
rsre
port
edin
this
tabl
e.
34
Tabl
e4:
Self
-Rep
orte
dH
ealt
han
dC
laim
sin
12M
onth
ssi
nce
Enro
llmen
t,by
Tem
pora
rySu
bsid
ies
and
Ass
iste
dIn
tern
etR
egis
trat
ion
Tab
le 4
Of
any
type
Out
pati
ent
Inpa
tien
tC
hron
icO
f an
y ty
peO
utpa
tien
tIn
pati
ent
Chr
onic
Val
ue o
f cl
aim
sD
ays
to
firs
t cla
im(1
)(2
)(3
)(4
)(5
)(6
)(7
)(8
)(9
)(1
0)(1
1)
Full
subs
idy
3.23
70.
480
0.45
70.
151
0.17
13.
591
3.37
90.
212
0.18
50.
986
232.
397
[0.4
52]
[0.5
01]
[0.4
99]
[0.3
59]
[0.3
77]
[6.8
05]
[6.6
06]
[0.5
69]
[0.4
29]
[2.6
20]
[141
.277
]H
alf
subs
idy
3.24
40.
511
0.49
40.
219
0.23
14.
698
4.35
80.
340
0.29
21.
879
213.
850
[0.5
41]
[0.5
01]
[0.5
01]
[0.4
15]
[0.4
22]
[10.
326]
[10.
071]
[0.8
13]
[0.6
24]
[4.7
12]
[150
.106
]N
o su
bsid
y3.
099
0.62
20.
612
0.18
10.
272
6.16
75.
906
0.26
20.
339
1.63
717
6.25
9[0
.538
][0
.486
][0
.489
][0
.386
][0
.446
][9
.714
][9
.340
][0
.696
][0
.600
][4
.064
][1
54.9
10]
Ass
iste
d in
tern
et r
egis
trat
ion
3.21
70.
525
0.50
30.
169
0.22
64.
253
3.99
60.
257
0.25
91.
366
214.
831
[0.5
24]
[0.5
00]
[0.5
01]
[0.3
75]
[0.4
19]
[7.4
18]
[7.1
79]
[0.6
96]
[0.5
12]
[3.6
76]
[147
.572
]St
atus
quo
reg
istr
atio
n3.
149
0.55
50.
537
0.19
80.
214
5.17
74.
903
0.27
30.
267
1.51
920
0.82
6[0
.505
][0
.498
][0
.499
][0
.399
][0
.411
][1
0.13
6][9
.842
][0
.668
][0
.583
][3
.852
][1
50.7
90]
Full
subs
idy
= n
o su
bsid
y0.
016
0.04
00.
039
0.27
30.
082
0.02
50.
025
0.26
50.
036
0.09
50.
006
Hal
f su
bsid
y =
no
subs
idy
0.01
40.
156
0.17
60.
536
0.63
80.
362
0.31
80.
483
0.67
60.
625
0.10
7A
ssis
ted
inte
rnet
reg
istr
atio
n =
st
atus
quo
0.08
40.
450
0.41
20.
423
0.78
70.
116
0.10
70.
787
0.81
50.
576
0.26
8
Had
a c
laim
Tot
al #
of
visi
tsC
laim
sSe
lf-
repo
rted
he
alth
P-v
alue
of
test
of
hypo
thes
is
Not
e:Th
ista
ble
show
sm
ean
self
-rep
orte
dhe
alth
and
mea
ncl
aim
ssu
bmit
ted
inm
onth
s1
to12
sinc
eon
e’s
enro
llmen
tda
teby
tem
pora
rysu
bsid
ies
and
assi
sted
inte
rnet
regi
stra
tion
.M
eans
are
wei
ghte
dto
refle
ctth
ein
tend
edra
ndom
izat
ion.
Stan
dard
devi
atio
nsar
ein
brac
kets
.Th
esa
mpl
eis
rest
rict
edto
hous
ehol
dsw
hoen
rolle
dw
ithi
na
year
sinc
eof
fer
and
had
cove
rage
for
atle
asto
nem
onth
over
the
sam
eti
me
peri
od.T
hesa
mpl
esi
zeis
749
hous
ehol
ds.I
nC
olum
n(1
),hi
gher
valu
esof
the
outc
ome
corr
espo
ndto
bett
erse
lf-r
epor
ted
heal
th.T
heva
lue
ofcl
aim
sin
Col
umn
(10)
isw
inso
rize
dat
the
99%
leve
land
only
refe
rsto
hosp
ital
clai
ms.
We
regr
ess
each
outc
ome
onin
dica
tor
vari
able
sfo
rtr
eatm
enta
ssig
nmen
t,an
indi
cato
rva
riab
lefo
rth
era
ndom
izat
ion
proc
edur
eus
edan
dan
indi
cato
rva
riab
lefo
rst
udy
loca
tion
(equ
atio
n(1
)).
All
regr
essi
ons
are
esti
mat
edby
OLS
and
wei
ghte
dto
refle
ctth
ein
tend
edra
ndom
izat
ion.
The
p-va
lues
repo
rted
are
from
ate
stof
the
diff
eren
cebe
twee
nth
eno
subs
idy
and
full
subs
idy
trea
tmen
ts(β
2=
0),b
etw
een
the
nosu
bsid
yan
dha
lfsu
bsid
ytr
eatm
ents
(β3=
0)an
dbe
twee
nth
est
atus
quo
and
assi
sted
inte
rnet
regi
stra
tion
trea
tmen
ts(β
4=
0).
All
regr
essi
ons
are
esti
mat
edby
OLS
and
wei
ghte
dto
refle
ctth
ein
tend
edra
ndom
izat
ion.
App
endi
xTa
ble
7pr
ovid
esth
ere
gres
sion
esti
mat
esbe
hind
the
num
bers
repo
rted
inth
ista
ble.
35
Tabl
e5:
Year
1C
laim
sby
Ret
enti
onin
Year
2,by
Tem
pora
rySu
bsid
ies
Tab
le 4
Of
any
type
Out
pati
ent
Inpa
tien
tC
hron
icO
f an
y ty
peO
utpa
tien
tIn
pati
ent
Chr
onic
Val
ue o
f cl
aim
sD
ays
to
firs
t cla
im(1
)(2
)(3
)(4
)(5
)(6
)(7
)(8
)(9
)(1
0)(1
1)
Dro
pout
s3.
198
0.41
40.
387
0.12
50.
154
2.55
32.
385
0.16
80.
154
0.75
525
1.25
1[0
.441
][0
.494
][0
.488
][0
.332
][0
.362
][4
.907
][4
.751
][0
.494
][0
.362
][2
.333
][1
34.7
32]
Stay
ers
3.31
30.
611
0.59
70.
204
0.20
55.
657
5.35
60.
301
0.24
61.
447
194.
896
[0.4
67]
[0.4
90]
[0.4
93]
[0.4
05]
[0.4
06]
[9.2
10]
[8.9
62]
[0.6
89]
[0.5
35]
[3.0
77]
[147
.187
]
Dro
pout
s =
sta
yers
0.06
80.
005
0.00
20.
146
0.33
60.
002
0.00
20.
134
0.14
00.
080
0.00
4
Dro
pout
s3.
248
0.41
30.
394
0.21
20.
155
2.77
32.
447
0.32
60.
192
1.60
024
0.24
0[0
.548
][0
.494
][0
.490
][0
.410
][0
.364
][5
.618
][5
.225
][0
.792
][0
.481
][4
.225
][1
47.4
98]
Stay
ers
3.23
80.
674
0.66
20.
231
0.35
77.
922
7.55
80.
364
0.45
92.
345
169.
670
[0.5
31]
[0.4
71]
[0.4
76]
[0.4
24]
[0.4
82]
[14.
737]
[14.
513]
[0.8
50]
[0.7
84]
[5.4
23]
[144
.754
]
Dro
pout
s =
sta
yers
0.91
40.
004
0.00
30.
778
0.00
60.
001
0.00
10.
793
0.01
20.
357
0.00
9
Dro
pout
s3.
169
0.53
00.
520
0.24
80.
234
3.63
33.
341
0.29
20.
327
1.84
119
7.47
2[0
.507
][0
.503
][0
.503
][0
.435
][0
.426
][5
.471
][5
.108
][0
.548
][0
.643
][4
.231
][1
57.0
35]
Stay
ers
3.05
50.
681
0.67
10.
138
0.29
67.
789
7.54
70.
242
0.34
61.
508
162.
689
[0.5
55]
[0.4
68]
[0.4
72]
[0.3
47]
[0.4
59]
[11.
377]
[10.
955]
[0.7
78]
[0.5
74]
[3.9
68]
[152
.736
]
Dro
pout
s =
sta
yers
0.18
80.
065
0.06
50.
126
0.43
90.
003
0.00
20.
636
0.87
30.
611
0.18
0
Pane
l C: N
o su
bsid
y
P-va
lue
of te
st o
f hy
poth
esis
P-va
lue
of te
st o
f hy
poth
esis
Self
-re
port
ed
heal
th
Had
a c
laim
Tot
al #
of
visi
tsC
laim
s
Pane
l A: F
ull s
ubsi
dy
P-va
lue
of te
st o
f hy
poth
esis
Pane
l B: H
alf
subs
idy
Not
e:T
his
tabl
esh
ows
mea
nse
lf-r
epor
ted
heal
than
dcl
aim
sin
the
first
year
sinc
een
rollm
ent,
sepa
rate
lyby
tem
pora
rysu
bsid
ies
and
byw
heth
erho
useh
olds
kept
ordr
oppe
dco
vera
geat
mon
th15
sinc
eof
fer
date
.Mea
nsar
ew
eigh
ted
tore
flect
the
inte
nded
rand
omiz
atio
n.St
anda
rdde
viat
ions
are
inbr
acke
ts.T
hesa
mpl
eis
rest
rict
edto
hous
ehol
dsw
hoen
rolle
dw
ithi
na
year
sinc
eof
fer
and
paid
for
atle
ast
one
mon
thov
erth
esa
me
tim
epe
riod
.Th
esa
mpl
esi
zeis
749
hous
ehol
ds.
InC
olum
n(1
),hi
gher
valu
esof
the
outc
ome
corr
espo
ndto
bett
erse
lf-r
epor
ted
heal
th.
The
valu
eof
clai
ms
inC
olum
n(1
0)is
win
sori
zed
atth
e99
%le
vela
ndon
lyre
fers
toho
spit
alcl
aim
s.Th
ep-
valu
esar
efr
oman
inte
ract
edsp
ecifi
cati
onw
here
the
outc
ome
isre
gres
sed
onin
dica
tor
vari
able
sfo
rtr
eatm
ent
assi
gnm
ent
inte
ract
edw
ith
anin
dica
tor
vari
able
for
whe
ther
the
hous
ehol
dha
sco
vera
gein
mon
th15
,an
indi
cato
rva
riab
lefo
rth
era
ndom
izat
ion
proc
edur
eus
edan
dan
indi
cato
rva
riab
lefo
rst
udy
loca
tion
.All
regr
essi
ons
are
esti
mat
edby
OLS
and
wei
ghte
dto
refle
ctth
ein
tend
edra
ndom
izat
ion.
App
endi
xTa
ble
11pr
ovid
esth
ere
gres
sion
esti
mat
esbe
hind
the
num
bers
repo
rted
inth
ista
ble.
36
Tabl
e6:
Expe
ndit
ures
and
Rev
enue
s,by
Tem
pora
rySu
bsid
ies
Cov
erag
eR
even
ues
Cla
ims
expe
ndit
ures
Net
rev
enue
sN
et r
even
ues
incl
udin
g ca
pita
tion
Rev
enue
sC
laim
s ex
pend
itur
esN
et r
even
ues
Net
rev
enue
s in
clud
ing
capi
tati
on
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Full
sub
sidy
2.88
70.
004
0.03
1-0
.027
-0.0
500.
001
0.00
7-0
.006
-0.0
12[4
.899
][0
.018
][0
.320
][0
.321
][0
.321
][0
.010
][0
.158
][0
.158
][0
.159
]H
alf
subs
idy
1.09
60.
042
0.19
4-0
.151
-0.1
700.
004
0.01
8-0
.013
-0.0
15[3
.013
][0
.041
][1
.571
][1
.567
][1
.567
][0
.018
][0
.478
][0
.475
][0
.476
]N
o su
bsid
y0.
323
0.07
50.
183
-0.1
08-0
.125
0.00
20.
005
-0.0
03-0
.003
[1.5
35]
[0.0
38]
[1.2
96]
[1.2
93]
[1.2
93]
[0.0
14]
[0.2
15]
[0.2
13]
[0.2
13]
Obs
erva
tion
s59
9655
5855
5855
5855
5871
952
7195
271
952
7195
2
Full
sub
sidy
= n
o su
bsid
y0.
000
0.00
00.
026
0.17
00.
198
0.00
20.
907
0.88
90.
253
Hal
f su
bsid
y =
no
subs
idy
0.00
00.
000
0.94
70.
776
0.75
50.
000
0.03
30.
073
0.04
2H
alf
subs
idy
= f
ull s
ubsi
dy0.
000
0.00
00.
012
0.05
00.
058
0.00
00.
114
0.25
20.
543
Full
sub
sidy
0.95
20.
067
0.09
3-0
.026
-0.0
470.
009
0.01
1-0
.003
-0.0
05[2
.317
][0
.054
][0
.691
][0
.692
][0
.692
][0
.029
][0
.240
][0
.239
][0
.239
]H
alf
subs
idy
0.59
30.
071
0.18
7-0
.116
-0.1
350.
006
0.01
4-0
.008
-0.0
09[1
.911
][0
.053
][1
.169
][1
.169
][1
.169
][0
.025
][0
.322
][0
.320
][0
.320
]N
o su
bsid
y0.
451
0.06
80.
153
-0.0
85-0
.102
0.00
40.
009
-0.0
05-0
.006
[1.6
80]
[0.0
47]
[1.4
32]
[1.4
31]
[1.4
31]
[0.0
19]
[0.3
42]
[0.3
40]
[0.3
40]
Obs
erva
tion
s59
9635
6535
6535
6535
6547
968
4796
847
968
4796
8
Full
sub
sidy
= n
o su
bsid
y0.
000
0.61
60.
778
0.73
60.
805
0.00
00.
141
0.65
80.
412
Hal
f su
bsid
y =
no
subs
idy
0.03
40.
296
0.37
70.
410
0.39
20.
003
0.18
20.
334
0.29
3H
alf
subs
idy
= f
ull s
ubsi
dy0.
000
0.76
00.
198
0.20
90.
221
0.04
20.
829
0.50
50.
655
P-v
alue
of
test
of
hypo
thes
is
P-v
alue
of
test
of
hypo
thes
is
Tab
le 6
Per
cov
ered
hou
seho
ld-m
onth
Per
hou
seho
ld-m
onth
Pan
el A
: Mon
ths
1 to
12
sinc
e of
fer
date
Pan
el B
: Mon
ths
13 to
20
sinc
e of
fer
date
Not
e:Th
ista
ble
show
sm
ean
reve
nues
and
expe
ndit
ures
byte
mpo
rary
subs
idie
s,fo
rmon
ths
1to
12fr
omof
fer(
Pane
lA)a
ndfo
rmon
ths
13to
20fr
omof
fer
(Pan
elB)
.Mea
nsar
ew
eigh
ted
tore
flect
the
inte
nded
rand
omiz
atio
n.St
anda
rdde
viat
ions
are
inbr
acke
ts.C
olum
n(1
)rep
orts
mea
nnu
mbe
rofm
onth
sw
ith
insu
ranc
eco
vera
ge.O
bser
vati
ons
are
atth
eho
useh
old
leve
l.C
olum
ns(2
)to
(5)a
nd(6
)to
(9)s
how
mea
nre
venu
es(p
rem
ium
spa
idby
enro
llees
),ex
pend
itur
es(t
otal
valu
eof
clai
ms)
,net
reve
nues
,and
netr
even
ues
incl
udin
gca
pita
tion
paym
ents
,in
mill
ions
IDR
for
hous
ehol
d-m
onth
sin
whi
chho
useh
olds
had
cove
rage
and
for
allh
ouse
hold
-mon
ths.
Obs
erva
tion
sar
eat
the
hous
ehol
d-m
onth
leve
l.T
heva
lue
ofcl
aim
sin
Col
umns
(5)
and
(9)i
sw
inso
rize
dat
the
99%
leve
land
only
refe
rsto
hosp
ital
clai
ms.
The
p-va
lues
are
from
regr
essi
ons
ofea
chou
tcom
eon
indi
cato
rva
riab
les
for
trea
tmen
tass
ignm
ent,
anin
dica
tor
vari
able
for
the
rand
omiz
atio
npr
oced
ure
used
and
anin
dica
tor
vari
able
for
stud
ylo
cati
on(e
quat
ion
(1))
.In
Col
umn
(1)s
tand
ard
erro
rsar
ero
bust
,whi
lein
Col
umns
(2)t
o(9
)sta
ndar
der
rors
are
clus
tere
dat
the
hous
ehol
dle
vel.
The
p-va
lues
repo
rted
are
from
ate
stof
the
diff
eren
cebe
twee
nth
eno
subs
idy
and
full
subs
idy
trea
tmen
ts(β
2=
0),b
etw
een
the
nosu
bsid
yan
dha
lfsu
bsid
ytr
eatm
ents
(β3=
0)an
dbe
twee
nth
eha
lfsu
bsid
yan
dfu
llsu
bsid
ytr
eatm
ents
(β1=
β2)
.A
llre
gres
sion
sar
ees
tim
ated
byO
LSan
dw
eigh
ted
tore
flect
the
inte
nded
rand
omiz
atio
n.A
ppen
dix
Tabl
e13
prov
ides
the
regr
essi
ones
tim
ates
behi
ndth
enu
mbe
rsre
port
edin
this
tabl
e.
37